What can be done with the Semantic Web?An Overview of Watson-based Applications?

Mathieu d’Aquin, Marta Sabou, Enrico Motta, Sofia Angeletou, LaurianGridinoc, Vanessa Lopez, and Fouad Zablith

Knowledge Media Institute (KMi)The Open University, Milton Keynes, United Kingdom

{m.daquin, r.m.sabou, e.motta, s.angeletou, l.gridinoc, v.lopez,

f.zablith}@open.ac.uk

Abstract. Thanks to the huge efforts deployed in the community forcreating, building and generating semantic information for the SemanticWeb, large amounts of machine processable knowledge are now openlyavailable. Watson is an infrastructure component for the Semantic Web,a gateway that provides the necessary functions to support applicationsin using the Semantic Web. In this paper, we describe a number of ap-plications relying on Watson, with the purpose of demonstrating whatcan be achieved with the Semantic Web nowadays and what sort of new,smart and useful features can be derived from the exploitation of thislarge, distributed and heterogeneous base of semantic information.

1 Introduction

It is now commonly admitted that, while it still needs to grow and evolve, theSemantic Web has already become a reality. Millions of RDF documents are nowpublished online, describing hundreds of millions of entities through billions ofstatements. The Semantic Web is now the biggest, most distributed and mostheterogeneous knowledge base that ever existed, and is very quickly evolving,as thousands of users constantly create new knowledge and update the existingone.

While it now appears clearly that the effort deployed by the community increating semantic information and making it available on the Web has beensuccessful, it seems to be a good time to consider the applications of this hugeinformation infrastructure the Semantic Web has become: What can be donewith it?

Watson [1] is a Semantic Web Gateway: it collects and indexes semantic in-formation on the Web, to provide a variety of access mechanisms for users andapplications. As such, it provides support for building a new kind of applica-tions taking benefit from the Semantic Web as it makes possible within theseapplications to dynamically find, explore and exploit online semantic content [2].? This work was funded by the Open Knowledge and NeOn projects sponsored by

the European Commission as part of the Information Society Technologies (IST)programme.

A number of applications relying on Watson have already been developed andprovide demonstrators of the possibilities offered by this approach of exploitingthe Semantic Web. In this paper, we describe these applications (at least, theones we know of, as Watson is an open service that anybody can use), withthe aim of providing an overview of the variety of tasks that can be achievednowadays with the Semantic Web.

2 Watson: the Semantic Web Gateway

The need of next generation Semantic Web applications for a new kind of infras-tructure motivated the development of Watson: a gateway that realizes a singleaccess point to the semantic information published online and provides efficientservices to support application developers in exploiting this large amount of dis-tributed and heterogeneous data.

2.1 Overview

Watson collects online semantic documents through a variety of crawlers, ex-ploring different sources, such as PingTheSemanticWeb.com. The particularityof these crawlers, compared with usual web crawlers, is that they consider se-mantic relations across documents in addition to classical hyperlinks. Also, whencollecting online semantic content they check for duplicates, copies, or prior ver-sions of the discovered documents.

Once collected, these documents are analyzed and indexed according to avariety of information concerning the content of the document, its complexity,quality and relations to other resources. This analysis step is core to Watson asit ensures that the key information is extracted, which can help applications inselecting, assessing, exploiting and combining these resources.

Fig. 1. Screenshots of the Watson Web interface (http://watson.kmi.open.ac.uk)

Finally, the goal of Watson is to provide efficient and adequate access to thecollected information for applications and, to some extent, human users. A webinterface allows users to search semantic content by keywords, to inspect them, toexplore semantic documents, and to query them using SPARQL (see Figure 1).

However, the important part of Watson is the services and API it provides tosupport the development of next generation Semantic Web applications. Indeed,Watson deploys a number of Web services and a corresponding API allowingapplications to:

– find Semantic Web documents through sophisticated keyword based search,allowing applications to specify queries according to a number of parameters(type of entities, level of matching of the keywords, etc.);

– retrieve metadata about these documents, e.g., size, language, label, logicalcomplexity, etc;

– find specific entities (classes, properties, individuals) within a document;– inspect the content of a document, i.e., the semantic description of the enti-

ties it contains;– apply SPARQL queries to Semantic Web documents.

The combination of mechanisms for searching semantic documents (keywordsearch), retrieving metadata about these documents and querying their content(e.g., through SPARQL) provides all the necessary elements for applications toselect and exploit online semantic resources. Moreover, the Watson web servicesand API are in constant evolution to support the requirements of novel applica-tions.

2.2 Differences with Other Semantic Web Search Engines

There are a number of similar systems to Watson, falling into the category ofSemantic Web search engines. However, Watson differs from these systems ina number of ways, the main one being that Watson is the only one to providethe necessary level of services for applications to dynamically exploit SemanticWeb data. For example, one of the most popular Semantic Web Search engine isSindice1. However, while Sindice indexes a very large amount of semantic data,it only provides a simple look-up service allowing applications/users to “locate”semantic documents. Therefore, it is still necessary to download and processthese documents locally to exploit them, which in many cases, is not feasible.The Swoogle system2 is closer to Watson. However, it does not provide some ofthe advanced search and exploration functions that are present in the WatsonAPIs (including the SPARQL querying facility). The Falcon-S3 Semantic Websearch engine has been focusing more on the user interface aspects, but nowprovides an initial API including a sub-set of the functions provided by Watson.

Another important aspect to consider is how open Semantic Web Searchengines are. Indeed, Watson is the only Semantic Web search engine to provideunlimited access to its functionalities. Sindice, Swoogle and Falcon-S are, on thecontrary, restricting the possibility they offer by limiting the number of queriesexecutable in a day or the number results for a given query.1 http://sindice.com/2 http://swoogle.umbc.edu/3 http://iws.seu.edu.cn/services/falcons/api/index.jsp

3 Services on Top of Watson

In this section, we look at a first category of applications of Watson: System thatmake use of Watson to provide additional features for users and developers.

3.1 Scarlet: Relation Discovery

Scarlet4 follows the paradigm of automatically selecting and exploring onlineontologies to discover relations between two given concepts. For example, whenrelating two concepts labeled Researcher and AcademicStaff, Scarlet 1) identifies(at run-time) online ontologies that can provide information about how these twoconcepts inter-relate and then 2) combines this information to infer their relation.We have investigated two increasingly sophisticated strategies to discover andexploit online ontologies for relation discovery. The first strategy, S1, derives arelation between two concepts if this relation is defined within a single onlineontology, e.g., stating that Researcher v AcademicStaff . The second strategy,S2 addresses those cases in which no single online ontology states the relationbetween the two concepts, by combining relevant information which is spreadover two or more ontologies - e.g., that Researcher v ResearchStaff in oneontology and that ResearchStaff v AcademicStaff in another. To supportthis functionality, Scarlet relies on Watson to access online ontologies.

Scarlet originates from the design of an ontology matcher that exploits the Se-mantic Web as a source of background knowledge to discover semantic relations(mappings) between the elements of two ontologies. This matcher was evaluatedin the context of aligning two large, real life thesauri: the UNs AGROVOCthesaurus (40K terms) and the United States National Agricultural Librarythesaurus NALT (65K terms) [3]. The matching process performed with bothstrategies resulted in several thousands mappings, using several hundreds onlineontologies, with an average precision of 70%.

3.2 The Watson Synonym Service

The Watson Synonym Service5 is a simple service that creates a base of termclusters, where the terms of a cluster are supposed to be associated to the samesense. It makes use of the information collected by Watson in the form of on-tologies to derive these clusters.

The basic algorithm to create term clusters is quite straightforward. Enti-ties in Semantic Web ontologies all possess one and only one identifier (in agiven namespace, e.g., we consider Person to be the identifier of http://www.example.org/onto#Person). They can also be associated to one or several la-bels, through the rdf:label property. Hence the algorithm simply assumes thata term t1 is a synonym of another term t2 if t1 and t2 are used either as labelor identifier of the same entity. Our synonym discovery offline algorithm thensimply iterates through all the entities in Watson’s ontologies to create clustersof terms that are used together in the identifiers or labels of entities.

4 http://scarlet.open.ac.uk/5 http://watson.kmi.open.ac.uk/API/term/synonyms

Of course, the quality of the results obtained with this method is not as goodas the one obtained with the complex and costly approaches that are employed tobuild systems like Wordnet. However, the advantage of this algorithm is that itsquality improves together with the growth of the Semantic Web, without requir-ing any additional effort for collecting the data. Quite a high number of good syn-onyms are found, like in the cluster {ending, death, termination, destruction}.In addition, this method does not only find synonyms in one language, but canprovide the equivalent terms in various languages, providing that multi-lingualontologies exist and cover these terms. It could be argued that these are notactually synonyms (but translations) and one of the possible extensions for thistool is to make use of the language information in the ontologies to distinguishthese cases.

4 Ontology Engineering with Watson

In this section, we look at applications that exploit online knowledge for thepurpose of creating new knowledge or enriching existing one.

4.1 The Watson Plugin for Knowledge Reuse

Ontology reuse is a complex process involving activities such as searching forrelevant ontologies for reuse, assessing the quality of the knowledge to reuse,selecting parts of it and, finally, integrating it in the current ontology project.As the Semantic Web provides more and more ontologies to reuse, there is anincreasing need for tools supporting these activities.

The Watson plugin6 aims to facilitate knowledge reuse by integrating thesearch capabilities of Watson within the environment of an ontology editor (theNeOn Toolkit). The resulting infrastructure allows the user to perform all thesteps necessary for large scale reuse of online knowledge within the same envi-ronment where this knowledge is processed and engineered.

In practice, the Watson plugin allows the ontology developer to find, in exist-ing online ontologies, descriptions of the entities present in the currently editedontology (i.e., the base ontology), to inspect these descriptions (the statementsattached to the entities) and to integrate these statements into the base ontology.For example, when extending the base ontology with statements about the classResearcher, the Watson plugin identifies, through Watson, existing ontologiesthat contain relevant statements such as:

– Researcher is a subclass of AcademicStaff– PhDStudent is a subclass of Researcher– Researcher is the domain of the property isAuthorOf

These statements can be used to extend the edited ontology, integrating themto ensure, for example, that the class Researcher becomes a subclass of a newlyintegrated class AcademicStaff.6 http://watson.kmi.open.ac.uk/editor plugins.html

4.2 Evolva: Ontology Evolution Using Background Knowledge

Ontologies form the backbone of Semantic Web enabled information systems.Today’s organizations generate huge amount of information daily, thus ontolo-gies need to be kept up to date in order to reflect the changes that affect thelife-cycle of such systems (e.g., changes in the underlying data sets, need for newfunctionalities, etc). This task, described as the “timely adaptation of an ontol-ogy to the arisen changes and the consistent management of these changes”, iscalled ontology evolution [4]. While it seems necessary to apply such a processconsistently for most of the ontology-based systems, it is often a time-consumingand knowledge intensive task, as it requires a knowledge engineer to identify theneed for change, perform appropriate changes on the base ontology and manageits various versions.

Evolva7 is an ontology evolution system starting from external data sources(text documents, folksonomies, databases, etc.) that form the most commonmean of storing data. First, a set of terms are extracted from these sources aspotentially relevant concepts/instances to add to the ontology, using commoninformation extraction methods. Evolva then makes use of Watson (throughthe intermediary of Scarlet) to find external sources of background knowledgeto establish relations between these terms and the knowledge already presentin the ontology, providing in this way the mean to integrate these new termsin the ontology. For this purpose, we devised a relation discovery process thatcombines various background knowledge sources with the goal of optimizingtime-performance and precision.

4.3 FLOR: FoLksonomy Ontology enRichment

Folksonomies, social tagging systems such as Flickr and Delicious, are at theforefront of the Web2.0 phenomenon as they allow users to tag, organize andshare a variety of information artifacts. The lightweight structures that emergefrom these tag spaces, only weakly support content retrieval and integration ap-plications since they are agnostic to the explicit semantics underlying the tagsand the relations among them. For example, a search for mammal ignores allresources that are not tagged with this exact word, even if they are tagged withspecific mammal names such as lion, cow, and cat. The objective of FLOR [7] isto attach formal semantics to tags, derived from online ontologies and make therelations between tags explicit (e.g., that mammal is a superclass of lion). Theenrichment algorithm that has been experimentally investigated builds on Wat-son: given a set of tags, the prototype identifies the ontological entities (classes,properties and individuals) that define the tags in their respective context. Ad-ditionally, it aims to identify formal relations between the tags (subsumption,disjointness and generic relations) utilizing Scarlet.

The experiments [8] have led to further insights into to nature of ontologieson the Semantic Web, from which we highlight two key ones. First, we found that

7 An overview of Evolva can be found in [5, 6].

online ontologies have a poor coverage of a variety of tag types denoting novelterminology scientific terms, multilingual terms and domain specific jargons.Secondly, we observed that online ontologies can reflect different views and whenused in combination can lead to inconsistencies in the derived structures.

5 End User Applications Relying on Watson

Finally, this section considers applications directly providing new features forusers by relying on online semantic content.

5.1 Wahoo/Gowgle: Query Expansion

Wahoo and Gowgle8 are 2 demonstrators, showing how Web ontologies can beused for a simple application to query expansion in a classical Web search engine.For example, when given a keyword like developer, such a tool could find outthat, in an ontology, there is a sub-class programmer of developer and couldtherefore suggest this term as a way to specify the query to the Web searchengine. Without Watson, this would require to integrate one or several ontologiesabout the domain of the queries and an infrastructure to store them, explorethem and query them. However, if the considered search engine is a general Websearch engine, such as Google or Yahoo, the domain of the queries cannot bepredicted a priori: the appropriate ontology can only be selected at run-time,depending on the query that is given. In addition, this application would requirea heavy infrastructure to be able to handle large ontologies and to query themefficiently. Gowgle and Wahoo rely on Semantic Web ontologies explored usingWatson instead.

The overall architecture of these applications is made of a Javascript/HTMLpage for entering the query and displaying the results, which communicates usingthe principles of AJAX with the Watson server. In the case of Gowgle, Googleis used as a the Web Search Engine and the Watson SOAP Web services areemployed for ontology exploration (http://watson.kmi.open.ac.uk/WS andAPI.html). In the case of Wahoo, Yahoo and the Watson REST API (http://watson.kmi.open.ac.uk/REST API.html) are used.

Both applications use Watson to exploit online ontologies, in order to suggestterms related to the query, that is, if the query contains the word developer : 1-to find ontologies somewhere talking about the concept of developer, 2- to findin these ontologies which entities correspond to developer and 3- to inspect therelations of these entities to find related terms.

5.2 PowerAqua: Question Answering

To some extent, PowerAqua9 can be seen as a straightforward human interfaceto any semantic document indexed by Watson. Using PowerAqua, a user can

8 http://watson.kmi.open.ac.uk/wahoo and http://watson.kmi.open.ac.uk/

gowgle9 http://poweraqua.open.ac.uk/

simply ask a question, like “Who are the members of the rock band Nirvana?”and obtain an answer, in this case in the form of a list of musicians (KurtCobain, Dave Grohl, Krist Novoselic and other former members of the group).The main strength of PowerAqua resides in the fact that this answer is deriveddynamically from the relevant datasets available on the Semantic Web. Notethat even if PowerAqua is meant to be an end-user application, lots of effort arestill required on interaction and user interface issues.

Without going into too many details, PowerAqua first uses a Gate-based [9]linguistic component to transform a question into a set of possible “query triples”,such as <person/organization, members, rock band Nirvana>. The next stepconsists then in locating, thanks to Watson, online semantic documents describ-ing entities that correspond to the terms of the query triples, locating for examplean individual called Nirvana in a dataset about music. During this step, Word-Net is used to augment the terms in the query triples with possible synonyms.Once a collection, usually rather large, of potential candidate ontologies is found,PowerAqua then employs a variety of heuristics and a powerful matching algo-rithm, PowerMap [10], to try and find answers from the collection of candidateontologies. In our example, the query triple shown above can be successfullymatched to the schema <Nirvana, has members, ?x:Musician>, which has beenfound in a music ontology on the Semantic Web. In more complex examples, ananswer may require integrating a number of statements. For instance, to answera query such as “Which Russian rivers flow to the Black Sea”, PowerAqua mayneed to find information about Russian rivers, information about rivers whichflow to the Black Sea and then combine the two. In general, several sources ofinformation, coming from various places on the Web, may provide overlappingor complementary answers. These are therefore ranked and merged according toPowerAqua’s confidence in their contribution to the final answer.

5.3 PowerMagpie: Semantic Browsing

PowerMagpie10 is a Semantic Web browser that makes use of openly availablesemantic data to support the interpretation process of the content of arbitraryweb pages. Unlike Magpie, which relied on a single ontology selected at designtime, PowerMagpie automatically, i.e., at run-time, identifies and uses relevantknowledge provided by multiple online ontologies. From a user perspective, Pow-erMagpie is an extension of a classical web browser and takes the form of avertical widget displayed on top of the currently browsed web page. This widgetprovides several functionalities that allow exploring the semantic informationrelevant to the current web page. In particular, it summarizes conceptual enti-ties relevant to the web page. Each of the entities can then be shown in the text,where the user may initialize different ways of exploring the information spacearound a particular entity. In addition, the semantic information discovered byPowerMagpie, which relates the text to online semantic resources, is injectedinto the web page as embedded annotations in RDFa. These annotations can

10 http://powermagpie.open.ac.uk

then be stored into a local knowledge base and act as an intermediary for theinteraction of different semantic-based systems.

5.4 Word Sense Disambiguation

Finally, we describe the work by Gracia et al. [11], which exploits large scalesemantics to tackle the task of word sense disambiguation (WSD). Specificallythey propose a novel, unsupervised, multi-ontology method which 1) relies ondynamically identified online ontologies as sources for candidate word senses and2) employs algorithms that combine information available both on the SemanticWeb and the web in order to integrate and select the right senses. The algorithmuses Watson to access online ontologies. Due to the rich Watson API, they canaccess all the important information without having to download the ontologies,thus providing an efficient and robust functionality.

The development and use of the WSD algorithm revealed that the SemanticWeb provides a good source of word senses that can complement traditionalresources such as WordNet. Also, because the extracted ontological informationis used as a basis for relatedness computation and not exploited through formalreasoning as in the case of ontology matching, the algorithm is less affected bythe quality of formal modeling than Scarlet. What affects the WSD method,however, is that most ontologies have a weak structure and as such provideinsufficient information to perform a satisfactory disambiguation.

6 Discussion: the Role of Watson

The availability of large amounts of semantic information is not sufficient inmaking the Semantic Web to achieve its full potential. Users should be providedwith smart, useful and efficient applications demonstrating the added value ofthe Semantic Web. The development of Watson is guided and informed by therequirements of such applications, which need an efficient and robust system toaccess the Semantic Web. In this paper, we provided an overview of a number ofapplications made possible by Watson, with the aim of showing what currentlycan be achieved with the Semantic Web and, hopefully, to provide inspirationfor developers to develop new applications.

In addition, we believe that these applications demonstrate the need for in-frastructure components like Watson, offering high level services to support thedynamic selection and exploitation of Semantic Web data. Indeed, a character-istic shared by the presented applications is that they all require to identifyand process semantic information at run-time, in an open domain. Consideringthis, usual approaches consisting in gathering relevant ontologies and datasetsin a local infrastructure are not applicable. It is necessary to gather the relevantknowledge dynamically, from various resources available on the Web.

The focus Watson put on providing a complete infrastructure for supportingapplication development is also essential to make these applications possible.Most of them would not be achievable using any other Semantic Web search en-gine, as it would require for them to download and process locally the semantic

documents. Apart from the obvious scalability issues this raises, it would alsobe needed for these applications to include an infrastructure capable of han-dling at run-time the heterogeneity of the semantic documents available online.Watson provides a common API to access pre-processed semantic documents inan homogeneous and efficient way. In addition, apart from this obvious reason,several of the presented applications use Watson instead of other similar systemseither because of the added functionalities it provides (e.g., Watson implementsadvanced ontology selection algorithms used by PowerMagpie) or because of itsrobustness (e.g. Scarlet switched from Swoogle to Watson for its ability to handleefficiently a large number of queries without restriction and without crashing).

However, one issue on which additional effort is required concerns the qualityof the delivered information. Indeed, as these applications relies on heterogeneousdata, coming from anywhere on the (Semantic) Web, it is essential to providesupport for assessing the quality of such data and to filter them accordingly.Watson already applies basic mechanisms to improve the quality of the results(e.g., filtering duplicates, extracting complexity information, etc.) A more flex-ible framework combining both automatic metrics for ontology evaluation anduser evaluation should be considered to make possible a more fine-grained, cus-tomizable quality ranking of semantic documents.

References

1. d’Aquin, M., Baldassarre, C., Gridinoc, L., Sabou, M., Angeletou, S., Motta, E.:Watson: Supporting Next Generation Semantic Web Applications. In: Proc. of theWWW/Internet conference. (2007)

2. dAquin, M., Motta, E., Sabou, M., Angeletou, S., Gridinoc, L., Lopez, V., Guidi,D.: Toward a New Generation of Semantic Web Applications. IEEE IntelligentSystems (2008)

3. Sabou, M., d’Aquin, M., Motta, E.: Exploring the Semantic Web as BackgroundKnowledge for Ontology Matching. Journal of Data Semantics (2008)

4. Haase, P., Stojanovic, L.: Consistent evolution of owl ontologies. European Se-mantic Web Conference (ESWC) (2005)

5. Zablith, F.: Dynamic ontology evolution. ISWC Doctoral Consortium (2008)6. Zablith, F., Sabou, M., d’Aquin, M., Motta, E.: Using Background Knowledge for

Ontology Evolution. In: International workshop on ontology dynamics. (2008)7. Angeletou, S., Sabou, M., Motta, E.: Semantically enriching folksonomies with

FLOR. In: Workshop Collective Intelligence & the Semantic Web. (2008)8. Angeletou, S., Sabou, M., Specia, L., Motta, E.: Bridging the gap between folk-

sonomies and the semantic web: An experience report. In: Workshop Bridging theGap between Semantic Web and Web 2.0. (2007)

9. Cunningham, H., Maynard, D., Bontcheva, K., Tablan, V.: GATE: A frameworkand graphical development environment for robust NLP tools and applications.In: Proc. of the 40th Anniversary Meeting of the Association for ComputationalLinguistics. (2002)

10. Lopez, V., Sabou, M., Motta, E.: PowerMap: Mapping the Real Semantic Web onthe Fly. In: International Semantic Web Conference (ISWC). (2006)

11. Gracia, J., Trillo, R., Espinoza, M., Mena, E.: Querying the Web: A MultiontologyDisambiguation Method. In: Proc. of the Sixth International Conference on WebEngineering (ICWE). (2006)